Modelpublic

Mihaj/nb-whisper-medium-karelian-CodeSwitching

sourceHugging Faceapache-2.0updated 2y agoView on Hugging Face

0likes4downloads

Model Card

nb-whisper-medium-karelian-CodeSwitching

This model is a fine-tuned version of NbAiLab/nb-whisper-medium-verbatim on the None dataset. It achieves the following results on the evaluation set:

—Loss: 0.5439
—Wer: 0.2585
—Cer: 0.0714

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

—learning_rate: 1e-05
—trainbatchsize: 16
—evalbatchsize: 16
—seed: 42
—distributed_type: multi-GPU
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 32
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: linear
—lrschedulerwarmup_steps: 500
—training_steps: 10000
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.2467	1.1351	500	0.5664	0.3488	0.0895
0.0718	2.2701	1000	0.5562	0.3166	0.0819
0.0513	3.4052	1500	0.5366	0.2997	0.0798
0.0377	4.5403	2000	0.5430	0.2815	0.0730
0.0339	5.6754	2500	0.5444	0.2906	0.0755
0.0263	6.8104	3000	0.5439	0.2757	0.0735
0.0182	7.9455	3500	0.5474	0.2754	0.0741
0.0141	9.0806	4000	0.5625	0.2808	0.0758
0.0117	10.2157	4500	0.5537	0.2662	0.0716
0.0122	11.3507	5000	0.5610	0.2703	0.0726
0.0118	12.4858	5500	0.5557	0.2686	0.0720
0.0075	13.6209	6000	0.5522	0.2673	0.0711
0.0069	14.7560	6500	0.5576	0.2764	0.0745
0.0072	15.8910	7000	0.5562	0.2676	0.0705
0.0085	17.0261	7500	0.5474	0.2713	0.0868
0.0041	18.1612	8000	0.5493	0.2639	0.0716
0.0041	19.2963	8500	0.5493	0.2612	0.0712
0.0041	20.4313	9000	0.5449	0.2554	0.0699
0.004	21.5664	9500	0.5444	0.2591	0.0708
0.0028	22.7015	10000	0.5439	0.2585	0.0714

Framework versions

—Transformers 4.41.2
—Pytorch 2.3.1+cu121
—Datasets 2.20.0
—Tokenizers 0.19.1